CN106204557A - A kind of extracting method of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere - Google Patents

Abstract

The extracting method of a kind of non-complete data symmetrical feature estimated with M based on extension Gaussian sphere, belongs to computer vision field.The present invention comprises the following steps: 1) uses method based on spatial digitizer to be scanned defect faceform, obtains initial mirror data.2) use the topological structure formatted based on space grating to set up spatial topotaxy between points, thus remove noise spot cloud, carry out K neighborhood search, then vow that information sets up extension Gaussian sphere, to going out corresponding point by method.3) corresponding point are slightly alignd based on main shaft.4) combine M and estimate the alignment of ICP algorithm essence.5) corresponding point Point Set least square fitting calculates centrosymmetry face.Extension Gaussian sphere is combined by the present invention with the advantage of the ICP algorithm of improvement so that feature extraction is more accurate, can meet the complex conditions of more feature extraction.

Description

A kind of extraction of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere Method

Technical field

The present invention relates to the symmetrical feature extracting method of a kind of non-complete data, particularly to one based on extension Gaussian sphere With the feature extracting method of the non-complete data of M estimation fusion, belong to computer vision field.

Background technology

Universal due to computer graphical scanning device and geometric modeling software, and the carrying of graphics process hardware cost performance Height, increasing threedimensional model has obtained applying widely in every field, the therefore feature extracting method of threedimensional model Become key.For the extraction of model, being first the pretreatment that model carries out coordinate, the more of employing is that main constituent divides Analysis method (Principal component analysis is called for short PCA), traditional PCA algorithm meeting in the retrieval of threedimensional model Arising a problem that, if carrying out, with PCA, the model that pretreatment tri patch describes, processing the PCA framework that obtains can be with The result of model trigonometric ratio and different.Novel PCA algorithm constantly occurred later, one of which for tringle coal, The area using each dough sheet is adjusted as weights, thus is solved above-mentioned problem.

Yang Qiong, Ding Xiaoqing et al. have also carried out certain improvement to PCA algorithm in 2003, and this algorithm is firstly introduced into mirror image and becomes Change, generate mirror image sample, then according to Parity-decomposition principle, generate the odd, even balanced sample of mirror image, and carry out K-L expansion respectively, Extract mirror image even odd symmetry KL characteristic component.Finally, according to even odd symmetry KL characteristic component shared energy proportion in face Difference and different sensitivitys that visual angle, rotation, illumination etc. are disturbed carry out feature selection.The outstanding advantage of this algorithm exists In significantly improving recognition performance, but, this algorithm is not particularly suited for the face organ of stronger asymmetry or bigger and positions by mistake The situation of difference.

Liyan Zhang, AnshumanRazdan et al. propose a fast automatic extraction face feature for 2005 Method, shooting 3D facial triagnle grid is as input, and first this method can automatically extract the bilateral symmetry plane of facial surface, Then the cross point of symmetrical plane and facial surface is calculated.Use average curvature figure and the symmetrical profiles curve of facial surface simultaneously Curvature chart, auto extractive is in three basic points of the nose of symmetrical profiles.These three basic point may determine that a face Intrinsic coordinate system (FICS).

Benz et al. and Hartmann et al. proposed one at 2005 and 2007 respectively and utilizes iterative closest point excellent The method changing (iterative closest point is called for short ICP), first, uses three and manually selects a little, when obtaining model During median surface facial planes, can obtain needing the data of mirror image by the information in intact district, main shaft side, then will with ICP algorithm Mirror image data carries out record with face information, and middle facial plane is defined as the best-fitting plane of point set, divides equally original Point is with the distance between mirror point.

Horn just proposed extension Gaussian image (extended Gaussian image is called for short EGI), Gaussian sphere in the early time It it is the funtion unit ball represented by a 3D shape.The method that Horn uses is each patch grids of threedimensional model Normal direction and area are mapped to Gaussian sphere, the normal direction of the corresponding mapped dough sheet in the direction of the centre of sphere of unit ball to spherical Map point, from And obtain the extension rectangular histogram of object.But extension Gaussian image has individual defect, for the threedimensional model of concave surface, its Gauss map Picture is not unique.

Therefore, it can utilize extension Gaussian sphere to combine with the ICP algorithm of M estimation fusion, come based on symmetrical feature Non-complete data carries out feature extraction.

Summary of the invention

The purpose of the present invention is to propose to the symmetrical feature of a kind of non-complete data newly estimated with M based on extension Gaussian sphere Extracting method, combines the advantage of the extension Gaussian sphere ICP algorithm with improving.

The present invention is achieved through the following technical solutions, a kind of non-complete data pair estimated with M based on extension Gaussian sphere Claim the extracting method of feature, comprise the steps:

1) defect face model is scanned, obtains initial mirror data；

2) set up extension Gaussian sphere and ask for corresponding point；

3) slightly align based on corresponding point；

4) corresponding point are carried out essence alignment by ICP algorithm to utilize M to estimate；

5) corresponding point Point Set least square fitting calculates centrosymmetry face.

Preferably, described step 1) defect face model is scanned, obtaining initial mirror data is: use three-dimensional to sweep The method retouching instrument, obtains facial contours and texture information.

Preferably, described step 2) set up extension Gauss map and ask for corresponding point and comprise the following steps:

(1) for the noise spot cloud removed in cloud data and set up spatial topotaxy, the point that scanning is obtained is needed Cloud data carry out pretreatment:

A. spatial topotaxy between points is set up initially with the topological structure formatted based on space grating；According to Cloud data is divided into equal-sized cubic space grid by the size of grid, and is returned by each point in cloud data Enter in corresponding cubic space grid；Grid division uses the structure of chained list to store cloud data after terminating；

B., on the basis of space structure rasterizing, the connectedness of grid is used to remove unnecessary noise spot cloud；

C. use a kind of k neighborhood fast search based on space separating strategy to point cloud searching；A given query point, searches Seek k the point closest with it, determine corresponding cubic space grid, compare whether counting more than k in grid, if It is the beeline just calculating point to six faces of sub-grid, grid carries out k neighborhood search, there to be k point to be in space End condition, records the sequence number of k point the grid that resets；

(2) law vector of curved surface is obtained

After data preprocessing, discrete point therein has curved surface characteristic, obtains the law vector i.e. characterization parameter of curved surface, Utilize least square fitting curved surface to be fitted nearest-neighbor point；First search for the k neighborhood point of P point and calculate its center of gravity O,

The method of least square fitting curved surface is vowed:

$f=\underset{i=1}{\overset{k}{\Σ}}\left|\right|(P_i-O)\·M\left|\right|$

Vector M is the normal vector of fitting surface, and i is interior some sequence numbers of k neighborhood of P point；

After trying to achieve law vector, its travel direction is adjusted, if measuring some P_i,P_j∈ S is 2 points that the distance on curved surface is close, The direction of normal of 2 should be consistent, the dot product m that two methods are vowed_i,m_j≈+1, otherwise represents that both are in opposite direction, m_iOr m_jShould be anti- To；

(3) law vector of cloud data is carried out unitization

Gauss Map is to carry out unitization by the law vector of cloud data on curve or curved surface, thus reflects curve or curved surface Geometrical property, and the starting point that method is vowed is moved on same end points, then to fall at radius be the unit circle of 1 to the end points that each method is vowed On, each method extremity point of curved surface then falls in unit sphere, the image that some subpoint on ball in method extremity is constituted, three-dimensional point The law vector of cloud is exactly that the method for this vertex neighborhood place fitting surface is vowed, if M is for putting cloud wherein, then its unit normal vector is P_M =(x, y, z), according to rectangular coordinate system with the conversion formula of spherical coordinate system be:

In formula: θ is P_MWith Z axis forward angle,For P_MWith X-axis forward angle；

The coordinate of Gaussian sphere is

(4) solve curvature of curved surface and calculate curvature a little

Utilizing analytic method to set up local coordinate system in a cloud, matching analytic surface in a coordinate system, by solving curved surface Curvature obtain curvature a little；After the law vector obtaining P calculated above, set up space coordinates, coordinate axes with a P for initial point For (l, m, n), using the direction of normal of a P as the direction of coordinate axes l, remaining coordinate axes is appointed in the incisal plane of a P and taken two Direction；

Curve surface definition is become vector form S (m, n)=[m n l (m, n)], the quadric parametric equation of minimum of matching ForWherein a=l_m, b=l_n, c=l_mm, d=l_mn, e=l_nn.Curved surface is at P point First-order partial derivative S_m=[1,0, l_m]^T, S_n=[0,1, l_n]^T, the per unit system at initial point is vowed and is

Curved surface second-order partial differential coefficient is S_mm=[0,0, l_mm]^T,S_mn=[0,0, l_mn]^T, S_nn=[0,0, l_nn]^T。

Defining and try to achieve curved surface first kind fundamental quantity A, B, C are as follows:

A=S_m ²=1+a², B=S_mS_n=ab, C=S_n ²=1+c²。

Defining and try to achieve curved surface Equations of The Second Kind fundamental quantity D, E, F are as follows:

$D=S_{mn}\·n=\frac{c}{\sqrt{1+a^2+b^2}},E=S_{mn}\·n=\frac{d}{\sqrt{1+a^2+b^2}},F=S_{nn}\·n=\frac{e}{\sqrt{1+a^2+b^2}}.$

The fundamental quantity of curved surface meets formulaWherein δ is the normal curvature of this point.

By first kind fundamental quantity A, B, C, Equations of The Second Kind fundamental quantity D, E, F substitute into above formula and obtain mean curvature H:

$H=\frac{{\mathrm{\δ}}_1+{\mathrm{\δ}}_2}{2}=\frac{AF-2BE+CD}{2(AC-B^2)}.$

(5) extension Gaussian sphere is set up

The curvature information of corresponding point is attached in Vector Message, forms extension Gaussian sphere；In extension Gaussian sphere, put cloud In any method vow starting point fall extension Gaussian sphere the centre of sphere on, end points falls on sphere, and each normal vector is with this point Curvature information, and realize a cloud space by the subscript index of point and extend the contacting of Gaussian sphere space；

(6) corresponding point in Searching point cloud

The purpose setting up extension Gaussian sphere is to quickly search for the corresponding point in two subject to registration somes clouds, by seeking Join template to determine overlapping region, it is achieved some cloud configuration based on extension Gaussian sphere, then to extension Gaussian sphere rotation transformation, with Complete the template matching of a cloud；Index by the subscript of a cloud and realize a cloud three dimensions and extend contacting of Gaussian sphere space, And find immediate point right by the curvature information of search corresponding point.

Preferably, described step 3) slightly it is aligned to based on corresponding point: utilize 3 alignment coordinate transformation methods, to a cloud number According to being initially registered, namely ask the transformation matrix between 2 clouds to make corresponding point slightly align.

Preferably, described step 4) M estimate ICP algorithm essence be aligned to:

The ICP method estimated based on M is to minimize lower array function E_r=∑_jρ(r_i(x_i,p)；σ), this function is residual error r_i (x_i, function p), this function has the effect of suppression exterior point so that exterior point works hardly, and (x, p)=t is then to make r

$\mathrm{\&rho;}(t_i,\mathrm{\&sigma;})=\left\{\begin{array}{cc}log(1+\frac{1}{2}{t_i}^2),& \left|t_i\right|<3\mathrm{\&sigma;}\\ 0,& \left|t_i\right|<3\mathrm{\&sigma;}\end{array}\right.$

Wherein σ is the standard deviation of residual error；Minimizing of E (r) is realized by conjugate gradient searching algorithm.

The invention has the beneficial effects as follows: by extension Gaussian sphere with the advantage of the ICP algorithm of improvement combines, thus propose A kind of new feature extracting method so that feature extraction is more accurate, can meet the complex conditions of more feature extraction.

Accompanying drawing explanation

The extracting method steps flow chart of the non-complete data symmetrical feature that Fig. 1 present invention estimates with M based on extension Gaussian sphere Figure；

The extension Gaussian sphere Establishing process figure of Fig. 2 present invention.

Detailed description of the invention

Below in conjunction with the accompanying drawings and the extraction side of symmetrical feature of non-complete data based on extension Gaussian sphere and M estimation fusion Method, is embodied as being further described to the present invention.

1) scanning defect face model

At present for gathering three-dimensional face data, use method based on spatial digitizer, it is possible to obtain the most accurate Facial contours is with texture information.

2) set up extension Gauss map, ask for corresponding point

(1) as in figure 2 it is shown, the method for building up of extension Gauss map is as follows: in order to remove the noise spot in cloud data Cloud also sets up spatial topotaxy, needs the cloud data obtaining scanning to process:

A. spatial topotaxy between points is set up initially with the topological structure formatted based on space grating.According to Cloud data is divided into equal-sized cubic space grid by the size of grid, and is returned by each point in cloud data Enter in corresponding cubic space grid.Grid division uses the structure of chained list to store cloud data after terminating.

B., on the basis of space structure rasterizing, the connectedness of grid is used to remove unnecessary noise spot cloud.

C. use a kind of k neighborhood fast search based on space separating strategy to point cloud searching.A given query point, searches Seek k the point closest with it, determine corresponding cubic space grid, compare whether counting more than k in grid, if It is the beeline just calculating point to six faces of sub-grid, grid carries out k neighborhood search, there to be k point to be in space End condition, records the sequence number of k point the grid that resets.

(2) law vector of curved surface is obtained

After data preprocessing, discrete point therein has curved surface characteristic, obtains the law vector i.e. characterization parameter of curved surface, Utilize least square fitting curved surface to be fitted nearest-neighbor point；First search for the k neighborhood point of P point and calculate its center of gravity O,

The method of least square fitting curved surface is vowed:

$f=\underset{i=1}{\overset{k}{\&Sigma;}}\left|\right|(P_i-O)\&CenterDot;M\left|\right|$

Vector M is the normal vector of fitting surface, and i is interior some sequence numbers of k neighborhood of P point；

After trying to achieve law vector, its travel direction is adjusted, if measuring some P_i,P_j∈ S is 2 points that the distance on curved surface is close, The direction of normal of 2 should be consistent, the dot product m that two methods are vowed_i,m_j≈+1, otherwise represents that both are in opposite direction, m_iOr m_jShould be anti- To；

(3) law vector of cloud data is carried out unitization

Gauss Map is to carry out unitization by the law vector of cloud data on curve or curved surface, thus reflects curve or curved surface Geometrical property, and the starting point that method is vowed is moved on same end points, then to fall at radius be the unit circle of 1 to the end points that each method is vowed On, each method extremity point of curved surface then falls in unit sphere, the image that some subpoint on ball in method extremity is constituted, three-dimensional point The law vector of cloud is exactly that the method for this vertex neighborhood place fitting surface is vowed, if M is for putting cloud wherein, then its unit normal vector is P_M =(x, y, z), according to rectangular coordinate system with the conversion formula of spherical coordinate system be:

In formula: θ is P_MWith Z axis forward angle,For P_MWith X-axis forward angle；

The coordinate of Gaussian sphere is

(4) solve curvature of curved surface and calculate curvature a little.Utilize analytic method to set up local coordinate system in a cloud, sitting Matching analytic surface in mark system, obtains curvature a little by the curvature solving curved surface.After the law vector obtaining P calculated above, Set up space coordinates with a P for initial point, coordinate axes be (l, m, n), using the direction of normal of a P as the direction of coordinate axes l, its Remaining coordinate axes is appointed in the incisal plane of a P and is taken both direction.

Curve surface definition is become vector form S (m, n)=[m n l (m, n)], the quadric parametric equation of minimum of matching ForWherein a=l_m, b=l_n, c=l_mm, d=l_mn, e=l_nn.Curved surface is at P point First-order partial derivative S_m=[1,0, l_m]^T, S_n=[0,1, l_n]^T, the per unit system at initial point is vowed and is Curved surface second-order partial differential coefficient is S_mm=[0,0, l_mm]^T,S_mn=[0,0, l_mn]^T, S_nn=[0,0, l_nn]^T。

Defining and try to achieve curved surface first kind fundamental quantity A, B, C are as follows:

A=S_m ²=1+a², B=S_mS_n=ab, C=S_n ²=1+c²。

Defining and try to achieve curved surface Equations of The Second Kind fundamental quantity D, E, F are as follows:

$D=S_{mn}\&CenterDot;n=\frac{c}{\sqrt{1+a^2+b^2}},E=S_{mn}\&CenterDot;n=\frac{d}{\sqrt{1+a^2+b^2}},F=S_{nn}\&CenterDot;n=\frac{e}{\sqrt{1+a^2+b^2}}.$

The fundamental quantity of curved surface meets formulaWherein δ is the normal curvature of this point.

By first kind fundamental quantity A, B, C, Equations of The Second Kind fundamental quantity D, E, F substitute into above formula and obtain mean curvature H:

$H=\frac{{\mathrm{\&delta;}}_1+{\mathrm{\&delta;}}_2}{2}=\frac{AF-2BE+CD}{2(AC-B^2)}.$

(5) extension Gaussian sphere is set up.The curvature information of corresponding point is attached in Vector Message, forms extension Gaussian sphere. In extension Gaussian sphere, the starting point that in some cloud, the method for any is vowed falls on the centre of sphere of extension Gaussian sphere, and end points falls on sphere, often Individual normal vector is all with the curvature information of this point, and it is empty with extension Gaussian sphere to realize a cloud space by the subscript index of point Between contact.

(6) corresponding point in Searching point cloud.The purpose setting up extension Gaussian sphere is to quickly search for 2 points subject to registration Corresponding point in cloud, determine overlapping region by seeking matching template, it is achieved some cloud configuration based on extension Gaussian sphere, then To extension Gaussian sphere rotation transformation, to complete the template matching of a cloud.Index by the subscript of a cloud and realize a cloud three dimensions With extension the contacting of Gaussian sphere space, and find immediate point right by the curvature information of search corresponding point.

3) slightly align based on corresponding point

After corresponding point have been searched for, utilize 3 alignment coordinate transformation methods, cloud data is initially registered, also It is exactly to ask the transformation matrix between 2 clouds to make corresponding point slightly align.Owing to algorithm exists local convergence sex chromosome mosaicism, therefore Need to combine ICP algorithm to align further.

4) M estimates the alignment of ICP algorithm essence

The ICP method estimated based on M is to minimize lower array function E_r=∑_jρ(r_i(x_i,p)；σ), this function is residual error r_i (x_i, function p), this function has the effect of suppression exterior point so that exterior point works hardly, and (x, p)=t is then to make r

$\mathrm{\&rho;}(t_i,\mathrm{\&sigma;})=\left\{\begin{array}{cc}log(1+\frac{1}{2}{t_i}^2),& \left|t_i\right|<3\mathrm{\&sigma;}\\ 0,& \left|t_i\right|<3\mathrm{\&sigma;}\end{array}\right.$

Wherein σ is the standard deviation of residual error；Minimizing of E (r) is realized by conjugate gradient searching algorithm.

(5) plane of symmetry is calculated

Utilize method of least square to calculate centrosymmetry face after the alignment of corresponding point essence.

Claims (5)

1. the extracting method of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere, it is characterised in that described Extracting method comprises the steps:

1) defect face model is scanned, obtains initial mirror data；

2) set up extension Gaussian sphere and ask for corresponding point；

3) slightly align based on corresponding point；

4) corresponding point are carried out essence alignment by ICP algorithm to utilize M to estimate；

5) corresponding point Point Set least square fitting calculates centrosymmetry face.

2. the extracting method of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere according to claim 1, its feature It is, described step 1) defect face model is scanned, obtaining initial mirror data is: use the side of spatial digitizer Method, obtains facial contours and texture information.

3. the extracting method of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere according to claim 1, its feature It is, described step 2) set up extension Gauss map and ask for corresponding point and comprise the following steps:

1) for the noise spot cloud removed in cloud data and set up spatial topotaxy, the cloud data that scanning is obtained is needed Carry out pretreatment:

A. spatial topotaxy between points is set up initially with the topological structure formatted based on space grating；According to grid Size cloud data is divided into equal-sized cubic space grid, and each point in cloud data is included into phase In the cubic space grid answered；Grid division uses the structure of chained list to store cloud data after terminating；

B., on the basis of space structure rasterizing, the connectedness of grid is used to remove unnecessary noise spot cloud；

C. use a kind of k neighborhood fast search based on space separating strategy to point cloud searching；A given query point, search with Its k closest point, determines corresponding cubic space grid, compares whether counting in grid is more than k, if just Calculate the some beeline to six faces of sub-grid, grid carries out k neighborhood search, in space, have k point for terminating Condition, records the sequence number of k point the grid that resets；

(2) law vector of curved surface is obtained

After data preprocessing, discrete point therein has curved surface characteristic, obtains the law vector i.e. characterization parameter of curved surface, utilizes Least square fitting curved surface is fitted nearest-neighbor point；First search for the k neighborhood point of P point and calculate its center of gravity O,

The method of least square fitting curved surface is vowed:

$f=\underset{i=1}{\overset{k}{\&Sigma;}}\left|\right|(P_i-O)\&CenterDot;M\left|\right|$

Wherein vector M is the normal vector of fitting surface, and i is interior some sequence numbers of k neighborhood of P point；

After trying to achieve law vector, its travel direction is adjusted, if measuring some P_i,P_j∈ S is 2 points that the distance on curved surface is close, 2 points Direction of normal should be consistent, the dot product m that two methods are vowed_i,m_j≈+1, otherwise represents that both are in opposite direction, m_iOr m_jShould be reverse；

(3) law vector of cloud data is carried out unitization

Gauss Map is to carry out unitization by the law vector of cloud data on curve or curved surface, thus reflects the several of curve or curved surface What characteristic, and the starting point that method is vowed is moved on same end points, the end points that each method is vowed then falls on the unit circle that radius is 1, Each method extremity point of curved surface then falls in unit sphere, the image that some subpoint on ball in method extremity is constituted, three-dimensional point cloud Law vector be exactly this vertex neighborhood place fitting surface method vow, if M for some cloud wherein, then its unit normal vector is P_M= (x, y, z), according to rectangular coordinate system with the conversion formula of spherical coordinate system be:

In formula: θ is P_MWith Z axis forward angle,For P_MWith X-axis forward angle；

The coordinate of Gaussian sphere is

(4) solve curvature of curved surface and calculate curvature a little

Utilizing analytic method to set up local coordinate system in a cloud, matching analytic surface in a coordinate system, by solving the song of curved surface Rate obtains curvature a little；After the law vector obtaining P calculated above, setting up space coordinates with a P for initial point, coordinate axes is (l, m, n), using the direction of normal of a P as the direction of coordinate axes l, remaining coordinate axes is appointed in the incisal plane of a P and is taken two sides To；

Curve surface definition becomes vector form S, and (m, n)=[m n l (m, n)], the quadric parametric equation of minimum of matching isWherein a=l_m, b=l_n, c=l_mm, d=l_mn, e=l_nn.Curved surface is the one of P point Rank partial derivative S_m=[1,0, l_m]^T, S_n=[0,1, l_n]^T, the per unit system at initial point is vowed and isBent Face second-order partial differential coefficient is S_mm=[0,0, l_mm]^T,S_mn=[0,0, l_mn]^T, S_nn=[0,0, l_nn]^T。

Defining and try to achieve curved surface first kind fundamental quantity A, B, C are as follows:

A=S_m ²=1+a², B=S_mS_n=ab, C=S_n ²=1+c²。

Defining and try to achieve curved surface Equations of The Second Kind fundamental quantity D, E, F are as follows:

$D=S_{mm}\&CenterDot;n=\frac{c}{\sqrt{1+a^2+b^2}},E=S_{mn}\&CenterDot;n=\frac{d}{\sqrt{1+a^2+b^2}},F=S_{nn}\&CenterDot;n=\frac{e}{\sqrt{1+a^2+b^2}}.$

The fundamental quantity of curved surface meets formulaWherein δ is the normal curvature of this point.

By first kind fundamental quantity A, B, C, Equations of The Second Kind fundamental quantity D, E, F substitute into above formula and obtain mean curvature H:

$H=\frac{{\mathrm{\&delta;}}_1+{\mathrm{\&delta;}}_2}{2}=\frac{AF-2BE+CD}{2(AC-B^2)}.$

(5) extension Gaussian sphere is set up

The curvature information of corresponding point is attached in Vector Message, forms extension Gaussian sphere；In extension Gaussian sphere, put Yun Zhongyi The starting point that the method for point is vowed falls on the centre of sphere of extension Gaussian sphere, and end points falls on sphere, and each normal vector is with the song of this point Rate information, and realize a cloud space by the subscript index of point and extend contacting of Gaussian sphere space；

(6) corresponding point in Searching point cloud

The purpose setting up extension Gaussian sphere is to quickly search for the corresponding point in two subject to registration somes clouds, by seeking coupling mould Plate determines overlapping region, it is achieved some cloud configuration based on extension Gaussian sphere, then to extension Gaussian sphere rotation transformation, to complete The template matching of some cloud；Index by the subscript of a cloud and realize a cloud three dimensions and extend contacting of Gaussian sphere space, and lead to The curvature information crossing search corresponding point finds immediate point right.

4. the extracting method of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere according to claim 1, its feature It is, described step 3) slightly it is aligned to based on corresponding point: utilize 3 alignment coordinate transformation methods, cloud data is carried out initially Registration, namely asks the transformation matrix between 2 clouds to make corresponding point slightly align.

5. the extracting method of the non-complete data symmetrical feature estimated with M based on extension Gaussian sphere according to claim 1, its feature Be, described step 4) M estimate ICP algorithm essence be aligned to:

The ICP method estimated based on M is to minimize lower array function E_r=∑_jρ(r_i(x_i,p)；σ), this function is residual error r_i(x_i,p) Function, this function have suppression exterior point effect so that exterior point works hardly, make r (x, p)=t, then

$\mathrm{\&rho;}(t_i,\mathrm{\&sigma;})=\left\{\begin{array}{cc}log(1+\frac{1}{2}{t_i}^2),& \left|t_i\right|<3\mathrm{\&sigma;}\\ 0,& \left|t_i\right|<3\mathrm{\&sigma;}\end{array}\right.$

Wherein σ is the standard deviation of residual error；Minimizing of E (r) is realized by conjugate gradient searching algorithm.